AI.png

On 23 June 2025, Microsoft introduced Mu, a compact language model designed to run directly on Windows devices. Built to enhance the new Settings agent, Mu is already available on preview builds for Copilot+ PCs, and it is provided free of charge.

 

What is Mu and why it matters

Mu is a 330-million parameter model developed by Microsoft’s research team. Unlike large cloud-based language models, Mu runs entirely on-device using the Neural Processing Unit (NPU). This allows it to respond to natural language commands like “turn on dark mode” or “adjust brightness” directly—without relying on cloud services or an internet connection.

The aim is to enable privacy-first, low-latency AI features within Windows 11, starting with the Settings app.

 

Key technical features of Mu

Mu is built using an encoder-decoder transformer architecture that processes inputs more efficiently than traditional decoder-only models. According to Microsoft, Mu delivers:

  • 47% faster first-token latency on Qualcomm NPUs
  • Up to 5× faster decoding speeds compared to standard decoder-only setups
  • Inference speeds over 100 tokens per second, reaching over 200 tokens/sec on devices like the Surface Laptop 7
  • Response times under 500 milliseconds

To further optimize Mu for Windows devices, Microsoft integrated features such as:

  • Shared input/output embedding layers
  • Rotary embeddings and grouped-query attention
  • Dual LayerNorm
  • 8-bit or 16-bit quantization for faster processing
  • Hardware-specific tuning for Intel, AMD, and Qualcomm NPUs

These improvements allow Mu to operate efficiently even on resource-constrained hardware.

 

How Mu powers the new Windows settings agent

Mu’s first use case is the Settings agent in Windows 11. This AI assistant helps users adjust system settings using natural language. For example, commands like:

  • “Turn on Bluetooth”
  • “Change wallpaper”
  • “Make the screen brighter”

These commands are executed directly by the system, without needing to navigate menus. Microsoft trained Mu on a custom dataset of 3.6 million examples, making it capable of handling hundreds of different setting adjustments with contextual awareness.

In cases where a command is unclear, the assistant will fall back on standard Settings search, providing helpful links instead of failing silently.

 

Current availability and future expansion

As of now, the Mu-powered Settings agent is available in preview for Windows Insiders running Copilot+ PCs with Qualcomm Snapdragon X processors. Microsoft confirmed that support for devices using Intel and AMD NPUs is on the way.

While no release date has been announced for general rollout, the Mu model’s efficient performance and low memory footprint suggest broader support could follow soon. Users in Malaysia and other regions will likely see these features as part of the next wave of on-device AI enhancements in Windows.

 

Would you use an AI agent to control your system settings instead of clicking through menus? Do you see on-device AI as the future of privacy-friendly computing? Share your thoughts with us. Stay tuned to TechNave.com for more updates.